---
source: https://iinventedlight.com/library/notes/2210-11416-flan-instruction-finetuning.md
title: "Scaling Instruction-Finetuned Language Models (Flan)"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, Scaling Instruction-Finetuned Language Models, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-05
as_of: 2026-11-05
confidence: medium
untested: "Results were not reproduced; only the arXiv abstract page was read."
category: ai-research
topic_area: "ai-research: instruction finetuning"
tags: llm, instruction-tuning, flan, chain-of-thought, arxiv
sources:
  - id: S1
    title: "Scaling Instruction-Finetuned Language Models"
    publisher: arXiv
    url: https://arxiv.org/abs/2210.11416
    accessed: 2026-11-05
    license: "CC BY 4.0"
---
# Scaling Instruction-Finetuned Language Models (Flan)

## Short answer
Finetuning a pretrained language model on many tasks written as instructions makes it better at new, unseen tasks. This paper scales that recipe along three axes: more tasks, bigger models, and chain-of-thought data. [S1]

## What they did
- Studied instruction finetuning while increasing the number of tasks, the model size, and adding chain-of-thought (step-by-step) examples. [S1]
- Applied it to several model families (PaLM, T5, U-PaLM) and tested zero-shot, few-shot, and chain-of-thought prompting. [S1]
- Flan-PaLM 540B, finetuned on about 1.8K tasks, beat the base PaLM 540B by 9.4% on average. [S1]
- Released Flan-T5 checkpoints, which perform well in few-shot settings even against much larger models. [S1]

## Why it matters
The authors present instruction finetuning as a general way to make pretrained models both stronger and easier to use. [S1]

## Limits
This note covers only the abstract. Benchmark numbers are as reported by the authors and were not re-run.

## Claims
- id: C1
  claim: The paper scales instruction finetuning by number of tasks, model size, and chain-of-thought data.
  source: S1
  evidence: "scaling the number of tasks, (2) scaling the model size"
- id: C2
  claim: Flan-PaLM 540B finetuned on 1.8K tasks outperforms PaLM 540B by 9.4% on average.
  source: S1
  evidence: "outperforms PALM 540B by a large margin (+9.4% on average)"
- id: C3
  claim: The authors released Flan-T5 checkpoints.
  source: S1
  evidence: "We also publicly release Flan-T5 checkpoints"
- id: C4
  claim: The authors call instruction finetuning a general method for improving pretrained models.
  source: S1
  evidence: "instruction finetuning is a general method for improving the performance"

## Sources
- [S1] Scaling Instruction-Finetuned Language Models. arXiv. https://arxiv.org/abs/2210.11416 . License: CC BY 4.0. Accessed 2026-11-05.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).